Emergent Behavioural Signatures in Large Language Models: A Cross-Task Study of Risk and Forecasting Behaviour

Recent advancements in large language models (LLMs) such as GPT-4, LLaMA, and Qwen2.5 have revealed capabilities extending beyond language generation to include complex reasoning and decision-making. This paper investigates whether LLMs exhibit consistent behavioural tendencies—comparable to human personality traits—when placed in structured decision-making scenarios. We conduct a two-pronged empirical study using (i) the Balloon Analogue Risk Task (BART), a psychological tool for assessing risk propensity, and (ii) a time-series forecasting task involving real-world FMCG sales data. Across both tasks, four state-of-the-art LLMs demonstrated stable and distinct behavioural profiles: models that acted conservatively in BART also generated cautious sales forecasts, while risk-taking models projected more aggressive outcomes. These patterns persisted across multiple runs and prompt variations, providing strong evidence that the observed behaviours are not artifacts of prompt engineering but rather emergent dispositions rooted in model architecture and training data. This work establishes a foundation for behavioural modelling in AI, with implications for building task-aligned foundation models that reflect consistent decision-making styles.